An Improved High-Resolution SST Climatology to ... - Semantic Scholar
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 4, JULY 2011
769
An Improved High-Resolution SST Climatology to Assess Cold Water Events off Florida Brian B. Barnes, Chuanmin Hu, and Frank Muller-Karger
Abstract—Cloud filters developed for high-resolution (1-km) Advanced Very High Resolution Radiometer (AVHRR) satellitederived sea surface temperature (SST) observations are generally inadequate to capture extreme cold events. Such events impacted shallow waters in Florida Bay and other coastal regions in January 2010 with fatal consequences for large numbers of corals and associated organisms. Raw AVHRR images were reprocessed to understand whether historical knowledge of daily and interannual SST variations could be used to derive a practical cloud-filtering technique. This approach, however, misidentified valid water temperature pixels in nearly 20% of 2703 images collected during the month of January for each year between 1995 and 2010. To create an improved SST climatology, this cloud-filtering method was combined with manually delineated overrides of falsely masked regions. During the January 2010 cold event, this climatology indicated negative SST anomalies of up to 11.6 ◦ C in the Big Bend region and 14 ◦ C in Florida Bay, with high spatial heterogeneity throughout. Our findings highlight the need for improved autonomous cloud-masking techniques to detect cold events in near real time. Index Terms—Ocean temperature, satellites, sea measurements.
I. I NTRODUCTION
A
SSESSMENT of the synoptic temperature variability that affects marine ecosystems requires accurate measurement of sea surface temperature (SST) from satellites, including observations from sensors such as the National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) and the National Aeronautics and Space Administration Moderate Resolution Imaging Spectroradiometer (MODIS). The accuracy of these observations under cloud-free conditions has been validated using in situ measurements [1]–[5], but cloud contamination leads to significant negative bias in SST estimates [6], [7]. Although many algorithms exist to filter clouds from unprocessed (level 0) data [8], [9] with varying success, no method successfully removes all clouds while retaining all valid data. A secondary (postprocessing) cloud filter has been proposed by Hu et al. [10] (hereafter termed “automated filter”) which is being implemented to autonomously remove cloudManuscript received October 28, 2010; revised December 21, 2010 and January 19, 2011; accepted January 25, 2011. Date of publication March 9, 2011; date of current version June 24, 2011. This work was supported by the National Aeronautics and Space Administration under Grants NNH10AN14I (through Florida Fish and Wildlife Conservation Commission Subcontract 09249) and NNX09AV24G to the University of South Florida. The authors are with the College of Marine Science, University of South Florida, St. Petersburg, FL 33701 USA (e-mail: [email protected]; [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2011.2111353
contaminated pixels in SST images at the University of South Florida (USF). Inputs for this filter are images that have been calibrated, navigated, cloud filtered, and processed to SST using TeraScan software (version 3.2; SeaSpace Corporation) as described by Hu et al. [10] (termed “raw” images). The automated filter relies on the temporal stability and climatological regularity of SST for a given location (image pixel)—removing pixels which differ by either 2◦ from a three-day moving median value or 4◦ from the weekly climatological mean [10]. These threshold values were determined using long-term in situ temperature data collected along the Florida Keys Reef Tract, and the results were validated for the same region. An extreme cold event affected South Florida and the Gulf of Mexico in January 2010. This event revealed that, for shallow waters close to land, valid SST data beyond the climatological threshold were mistakenly discarded as clouds by the automated filter. As the cloud filter was not designed to retain anomalies larger than 4 ◦ C, other cold events in the past might also have been improperly masked for nearshore waters. An unintended consequence of this artifact is the generation of a positively biased SST climatology. The problem of false cloud masking would thus be exacerbated in filtering of future images. All SST cloud-masking approaches seek to maximize cloud exclusion while retaining valid SST observations. The failure of the automated filter to detect the January 2010 cold event required a manual retrospective assessment of the filter performance and subsequent analysis of the climatology errors. Accurate SST assessment is crucial, as physiological deficiencies and even mortality have been reported in response to cold SST for a variety of resident organisms including corals [11], [12], manatees [13], [14], sea turtles [15], mangroves [16], [17], and fishes [16], [18]. Apart from being critical to ecosystem health overall, each of these species contributes economic benefits to the State of Florida. Knowledge of extent and severity of extreme-temperature events is required to effectively monitor this coastal ecosystem and may help direct real-time research or rescue efforts. Our objective was thus to develop an improved high-resolution SST climatology to study cold events around Florida. II. M ETHODS In response to the cold event in January 2010, existing SST climatology and individual cloud-filtered images for the month of January (1995–2010) from the USF AVHRR ground station and data archive were examined. We extracted all images collected for each January from 1995 to 2010 covering waters
1545-598X/$26.00 © 2011 IEEE
770
Fig. 1. AVHRR images from the NOAA-15 satellite collected on January 15, 2010, 10:37 GMT. (a) Raw image showing cold waters off Florida’s west coast around (yellow oval) Big Bend and (red oval) cloudinduced artifacts. (b) Automated postprocessing cloud filter (red oval) accurately masks cloud pixels but (yellow oval) falsely masks cold valid SST data. (c) (White) Manually created ROI over the raw image to define to which pixels the automated filter should not be applied. (d) Final image from the hybrid postprocessing filter (combining the automated filter and manual ROI delineation).
around Florida (24◦ to 31◦ N, 79◦ to 86◦ W), representing a database of 3200 satellite images. Passes where the swath measured less than approximately 25% of the coverage area were discarded, leaving 2703 images. Raw SST images processed with TeraScan to SST but not filtered using the secondary cloud-screening algorithm of Hu et al. [10] were first visually compared side by side to corresponding postprocessing filtered images [Fig. 1(a) and (b)]. Where it was apparent that the filtered image was falsely masking valid SST data, the raw image was set aside for reprocessing. This determination of valid data was based on integration of spatial and temporal information. Specifically, improper masking was documented only when the automated cloud filter removed a region-wide (approximately 200 km2 or more) feature which persisted in three or more consecutive images and displayed smooth spatial gradients consistent with that expected from the underlying bathymetry and water flow regime. The selected raw images were loaded into IDL/ENVI (version 4.5; Research Systems, Inc.) image processing software, and a region of interest (ROI) outlining the falsely masked pixels was manually created using ENVI’s ROI tool [Fig. 1(c)]. Raw data pixels covered by the ROI were passed through the automated filter to generate a new image. Those pixels outside the ROI were superimposed on the new image without modification [Fig. 1(d)]. In the creation of the ROIs, special attention was given to coastal locations as these are more likely to show faster and stronger temperature fluctuations in response to atmospheric weather patterns. In this way, the resulting filter (hereafter termed “hybrid filter”) retained the pixels deemed
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 4, JULY 2011
Fig. 2. Differences between 1995–2010 climatologies derived from the automated filter and hybrid filter. (a) January 1–7. (b) January 8–14. (c) January 15– 21. (d) January 22–28. Negative numbers indicate that the climatology from the hybrid filter has lower SST than the climatology from the automated filter. White is zero. Areas within the yellow and red outline boxes show zoomed-in images of the Florida Bay and southern Big Bend regions, respectively.
valid from both the raw and automatically filtered images, maximizing the amount of valid data in the image. Although this method might introduce a small degree of subjectivity, interpretation from different analyzers showed excellent consistency in identifying the misclassified pixels. Note that this manual effort was not to manually delineate clouded or mixed pixels but instead only to define a crude ROI with which to override the automated filter. Two sets of monthly and weekly climatologies were created for January 1995–2010. The first was constructed using only the images processed with the automated filter (i.e., including those with false cloud-masking artifacts). The second climatology used the hybrid-filtered images and those images in the series with no artifacts detected. The difference between the two weekly climatologies was calculated by subtracting the first (automated) from the second (hybrid; Fig. 2). Finally, the reprocessed images and new climatologies were used to assess the extent and severity of the January 2010 cold event in Florida. Daily averaged SST composites from January 10 to 14, 2010, were subtracted from the weekly climatology to create anomaly maps (Fig. 3). To evaluate the results of the hybrid filter, images from both filters were compared to in situ data collected from the archived National Data Buoy Center (NDBC) sea temperature data. Of the 66 NDBC stations that measured water temperature within the region studied, 26 included data that matched the AVHRR data in both time and space (Fig. 4). Data from both sources were used only if the sample times were within 0.5 h. This led to 10 578 matching pairs for the automated filter and 10 900 matching pairs for the hybrid filter. Linear regressions, root-mean-square (rms) difference, and bias calculations were performed using IDL (version 7.0; ITT Visual Information Solutions).
BARNES et al.: IMPROVED SST CLIMATOLOGY TO ASSESS COLD WATER EVENTS OFF FLORIDA
Fig. 3. Daily composite SST anomalies on (a) and (b) January 11, (c) and (d) January 12, and (e) and (f) January 13, 2010. Images (a), (c), and (e) were created using the automated filter with the original climatology, whereas images (b), (d), and (f) were generated with the current hybrid filter and improved climatology. White pixels represent clouds or invalid data.
Fig. 4.
Map of NDBC stations used and regions discussed in this letter.
III. R ESULTS Of the 2703 images analyzed, 498 images (18.4%) showed false cloud-masking artifacts from the automated filter. The largest portion of the images that required reprocessing were
771
from extreme cold events in 2001 and 2010. The automated filter regularly masked extremely cold (i.e., differing by > 4 ◦ C from climatology) but valid SST data in nearshore waters (see Fig. 1). The incorrect masking occurred around most of Florida’s coast, particularly in the Florida Bay and Big Bend regions. Several images also showed that the automated filter occasionally incorrectly masked warm-temperature data, particularly during warm spells in January of 1997 and 2002. The weekly climatologies created with the hybrid filter showed substantially different average SST (+1 ◦ C to −2 ◦ C) than the climatologies based on the automated cloud screening. The most extreme deviations were observed nearshore during the first and second weeks of January. In the first week (January 1 to 7), the original climatology showed overall higher average SST than the hybrid climatology over the entire west coast of Florida. In the Big Bend and Florida Bay regions, this new climatology was cooler by up to 2 ◦ C. The second week of January (days 8 to 14) showed temperatures approximately 0.5 ◦ C lower in the nearshore regions but up to 1 ◦ C higher in some areas of central western Florida and in Florida Bay [Fig. 2(b)]. Neither the third (January 15 to 21) nor the fourth (January 22 to 28) week showed any large differences between the two climatologies. The daily SST composites created with the two filters were compared to their corresponding filtered monthly climatologies. The original automated filter misidentified most pixels located in waters shallower than 20 m as clouds during the extreme cold event on January 11–13, 2010 (Fig. 3). The images processed with the hybrid filter, however, show that the daily averaged SST in Florida Bay during this span had negative anomalies of up to 14 ◦ C. The maximum negative anomaly for the Big Bend region was 11.6 ◦ C. The coldest temperatures occurred on January 11, 2010, with a slight temperature increase over the next two days (Fig. 3). Unfortunately, cloud cover on January 10 and 14 precluded region-wide examination of the days immediately preceding and following these extremetemperature days. The automatically filtered AVHRR data had a strong positive relationship with near-surface SST observations from NDBC buoys (r2 = 0.896, n = 10 578, linear regression slope = 0.923, and intercept = 1.753 ◦ C; Fig. 5). The rms difference was 0.927 ◦ C, and the bias was 0.097 ◦ C. The hybrid filter resulted in more matching pairs and also higher accuracy statistics. The NDBC and hybrid-filtered AVHRR data showed a strong positive relationship (r2 = 0.911, n = 10 900, slope = 0.972, intercept = 1.323 ◦ C, rms difference = 0.925 ◦ C, and bias = 0.084 ◦ C; Fig. 5). The 333 additional matching pairs (range: 6.5 ◦ C to 27.5 ◦ C) recovered with the hybrid filter were also strongly correlated (r2 = 0.979, slope = 1.005, intercept = −0.406 ◦ C, rms difference = 0.854 ◦ C, and bias = −0.322 ◦ C). These comparisons demonstrate the high quality of the AVHRR SST estimates derived using the standard algorithms, which were improperly masked under unusually cold or warm conditions. The automated filter incorrectly omitted matching pairs below 10 ◦ C and approximately 30% of those between 10 ◦ C and 15 ◦ C and also had more than 10% misclassifications for pixels matched at temperatures > 26 ◦ C (Fig. 6).
772
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 4, JULY 2011
Fig. 5. Comparison of temperature measurements from AVHRR and NDBC sources, taken at the same location and within 0.5 h. Blue points are those captured by AVHRR images processed with the automated filter. Orange points are the additional data from the hybrid filter. Data from January 10–13, 2010, are also denoted, most of which are missed by the automated filter.
Fig. 6. Percentage of matched AVHRR/NDBC data pairs captured by the hybrid filter that were missed by the automated filter, binned in 1 ◦ C intervals. The matched AVHRR/NDBC data pairs were taken at the same location and within 0.5 h.
IV. D ISCUSSION Nearly one in five images from January 1995–2010 was improved using the hybrid filter. Such misclassifications, however, may also occur in nonwinter months. For example, improper cloud masking may occur in the wake of a hurricane, in response to unseasonal upwelling of cold water [19], or under extremely calm and warm conditions. The primary misclassifications occurred north of the Florida Keys and in the Big Bend region—shallow waters heavily influenced by rapidly fluctuating air temperatures, insolation, and wind. Colder-thanexpected water was often seen also around river outflows, owing to the mixing of cold riverine waters with marine waters in these regions.
The difference between climatologies computed using the different filters was not large, primarily because the inputs for the two climatologies were identical in 81.6% of the images and most of the pixels in the reprocessed images. Furthermore, the automated filter had improperly masked both warm and cold waters at specific locations, and these errors cancelled to some degree in the climatology calculations. However, without the hybrid approach, it would be impossible to understand and quantify either positive or negative errors. The effect can be clearly visualized in Fig. 5. The images processed with the hybrid filter helped to extend the range of the observed temperatures to ∼ 6 ◦ C and improved the matchup statistics (Figs. 5 and 6), resulting in more data coverage on individual images. Validation against NDBC data also served to justify the manual override methodology, as the additional matching pairs showed high correlation and small error. There was a minor cold bias in these data (AVHRR data colder than NDBC data) in contrast to the slight warm bias of the entire data set. The cold bias is primarily driven by a cluster of additional matching pairs between approximately 8 ◦ C and 10 ◦ C, nearly all of which are from a single location (station 41112) during the 2010 cold event. Given the high spatial heterogeneity during this time near the St. Johns River plume, these cold-biased matches possibly are artifacts of the different spatial scales of measurement (point buoy versus 1-km2 pixel satellites), or they may indicate a slight calibration error at one NDBC station. The problem of false cloud masking of cold pixels is not unique to this region [9] or to the AVHRR sensor, as it is even present in the most sophisticated MODIS observations and cloud-screening algorithms [20]. Although the visual interpretation of thousands of images appears tedious, in practice, it is feasible and represents the best approach for climatology development. The work done here manually would nevertheless be difficult to implement in an autonomous process, yet it does serve as an example in developing the most accurate high-resolution AVHRR SST climatology for coastal regions in the global ocean that may experience extreme-temperature fluctuations. An autonomous climatology-based filter which allowed all anomalies observed in this study (up to 14 ◦ C) would clearly be insufficient to remove clouds from this system. Alternatively, lowering filter thresholds for nearshore waters (for example, by applying a bathymetry mask) or during winter months could potentially create artificial “fronts” in the climatologies at these spatial or temporal boundaries, which might further affect subsequent image processing. Given the strong correlation between the AVHRR and NDBC data, these in situ sensors could potentially be integrated into the cloud-filtering methodology. Such a system might allow autonomous overrides of falsely masked areas but would require a more extensive sensor network to be synoptically effective. At present, the possibility of extreme temperatures in coastal waters of Florida (and similar coastal waters in the global ocean) highlights the need for improved autonomous cloudmasking methods for high-resolution satellite imagery. Furthermore, for effective monitoring of environmental parameters, it is important to continue coordinated efforts to capture SST and
BARNES et al.: IMPROVED SST CLIMATOLOGY TO ASSESS COLD WATER EVENTS OFF FLORIDA
bottom temperatures from local in situ sensors. Such systems allow real-time validation of filtered satellite data and provide more thorough information of the thermal conditions throughout the water column. ACKNOWLEDGMENT The authors would like to thank the staff of the University of South Florida (USF) Institute for Marine Remote Sensing (IMaRS) for their acquisition and geonavigation of the Advanced Very High Resolution Radiometer database. The authors would also like to thank the anonymous reviewers and the many contributors within IMaRS and the USF Optical Oceanography Laboratory who helped to greatly improve this letter. The National Data Buoy Center historical temperature measurements are provided by the National Oceanic and Atmospheric Administration (available at www.ndbc.noaa.gov), and the Florida coastline layer (Fig. 4) is obtained from the Florida Fish and Wildlife Research Institute. R EFERENCES [1] A. E. Strong and E. P. McClain, “Improved ocean surface temperatures from space—Comparisons with drifting buoys,” Bull. Amer. Meteorol. Soc., vol. 65, no. 2, pp. 138–142, Feb. 1984. [2] A. F. Pearce, A. J. Prata, and C. R. Manning, “Comparison of NOAA/AVHRR-2 sea surface temperatures with surface measurements in coastal waters,” Int. J. Remote Sens., vol. 10, no. 1, pp. 37–52, Jan. 1989. [3] W. G. Pichel, “Operational production of multichannel sea surface temperatures from NOAA polar satellite AVHRR data,” Global Planet. Change, vol. 4, no. 1–3, pp. 173–177, Jul. 1991. [4] C. C. Walton, W. G. Pichel, J. F. Sapper, and D. A. May, “The development and operational application of nonlinear algorithms for the measurement of sea surface temperatures with the NOAA polar-orbiting environmental satellites,” J. Geophys. Res., vol. 103, pp. 27 999–28 012, 1998. [5] X. Li, W. G. Pichel, P. Clemente-Colón, V. Krasnopolsky, and J. Sapper, “Validation of coastal sea and lake surface temperature measurements derived from NOAA/AVHRR data,” Int. J. Remote Sens., vol. 22, no. 7, pp. 1285–1303, 2001.
773
[6] E. P. McClain, W. G. Pichel, C. C. Walton, Z. Ahmad, and J. Sutton, “Multi-channel improvements to satellite-derived global sea surface temperatures,” Adv. Space Res., vol. 2, no. 6, pp. 43–47, 1983. [7] A. K. Heidinger, V. R. Anne, and C. Dean, “Using MODIS to estimate cloud contamination of the AVHRR data record,” J. Atmos. Ocean. Technol., vol. 19, no. 5, pp. 586–601, May 2002. [8] R. W. Saunders and K. T. Kriebel, “An improved method for detecting clear sky and cloudy radiances from AVHRR data,” Int. J. Remote Sens., vol. 9, no. 1, pp. 123–150, Jan. 1988. [9] C. J. Merchant, A. R. Harris, E. Maturi, and S. Maccallum, “Probabilistic physically based cloud screening of satellite infrared imagery for operational sea surface temperature retrieval,” Q. J. R. Meteorol. Soc., vol. 131, pp. 2735–2755, 2005. [10] C. Hu, F. Muller-Karger, B. Murch, D. Myhre, J. Taylor, R. Luerssen, C. Moses, C. Zhang, L. Gramer, and J. Hendee, “Building an automated integrated observing system to detect sea surface temperature anomaly events in the Florida Keys,” IEEE Trans. Geosci. Remote Sens., vol. 47, pp. 2071–2084, Jul. 2009. [11] P. Jokiel and S. Coles, “Effects of temperature on the mortality and growth of Hawaiian reef corals,” Mar. Biol., vol. 43, no. 3, pp. 201–208, Sep. 1977. [12] A. E. Douglas, “Coral bleaching—How and why?” Mar. Pollut. Bull., vol. 46, no. 4, pp. 385–392, Apr. 2003. [13] A. B. Irvine, “Manatee metabolism and its influence on distribution in Florida,” Biol. Conserv., vol. 25, no. 4, pp. 315–334, Apr. 1983. [14] C. J. Deutsch, J. P. Reid, R. K. Bonde, D. E. Easton, H. I. Kochman, and T. J. O’Shea, “Seasonal movements, migratory behavior, and site fidelity of west Indian manatees along the Atlantic coast of the United States,” Wildlife Monogr., vol. 151, pp. 1–77, 2003. [15] B. E. Witherington and L. M. Ehrhart, “Hypothermic stunning and mortality of marine turtles in the Indian river lagoon system, Florida,” Copeia, vol. 1989, pp. 696–703, 1989. [16] M. Storey and E. W. Gudger, “Mortality of fishes due to cold at Sanibel Island, Florida 1886–1936,” Ecology, vol. 17, no. 4, pp. 640–648, Oct. 1936. [17] T. Savage, “Florida mangroves as shoreline stabilizers,” Florida Dept. Natural Resour., St. Petersburg, FL, Prof. Pap. No. 19, 1972, 46 pp. [18] Y. Sadovy and A.-M. Eklund, “Synopsis of Biological Data on the Nassau Grouper, Epinephelus Striatus (Bloch, 1792), and the Jewfish, E. Itajara (Lichtenstein, 1822),” Seattle, WA, U.S. Dept. Commerce, Nat. Ocean. Atmos. Admin., Nat. Mar. Fisheries Service, 65 pp., 1999. [19] N. D. Walker, R. R. Leben, and S. Balasubramanian, “Hurricane-forced upwelling and chlorophyll a enhancement within cold-core cyclones in the Gulf of Mexico,” Geophys. Res. Lett., vol. 32, p. L18 610, Sep. 2005. [20] S. A. Ackerman, K. I. Strabala, W. P. Menzel, R. A. Frey, C. C. Moeller, and L. E. Gumley, “Discriminating clear sky from clouds with MODIS,” J. Geophys. Res.—Atmos., vol. 103, no. D24, pp. 32 141–32 157, Dec. 1998.
769
An Improved High-Resolution SST Climatology to Assess Cold Water Events off Florida Brian B. Barnes, Chuanmin Hu, and Frank Muller-Karger
Abstract—Cloud filters developed for high-resolution (1-km) Advanced Very High Resolution Radiometer (AVHRR) satellitederived sea surface temperature (SST) observations are generally inadequate to capture extreme cold events. Such events impacted shallow waters in Florida Bay and other coastal regions in January 2010 with fatal consequences for large numbers of corals and associated organisms. Raw AVHRR images were reprocessed to understand whether historical knowledge of daily and interannual SST variations could be used to derive a practical cloud-filtering technique. This approach, however, misidentified valid water temperature pixels in nearly 20% of 2703 images collected during the month of January for each year between 1995 and 2010. To create an improved SST climatology, this cloud-filtering method was combined with manually delineated overrides of falsely masked regions. During the January 2010 cold event, this climatology indicated negative SST anomalies of up to 11.6 ◦ C in the Big Bend region and 14 ◦ C in Florida Bay, with high spatial heterogeneity throughout. Our findings highlight the need for improved autonomous cloud-masking techniques to detect cold events in near real time. Index Terms—Ocean temperature, satellites, sea measurements.
I. I NTRODUCTION
A
SSESSMENT of the synoptic temperature variability that affects marine ecosystems requires accurate measurement of sea surface temperature (SST) from satellites, including observations from sensors such as the National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) and the National Aeronautics and Space Administration Moderate Resolution Imaging Spectroradiometer (MODIS). The accuracy of these observations under cloud-free conditions has been validated using in situ measurements [1]–[5], but cloud contamination leads to significant negative bias in SST estimates [6], [7]. Although many algorithms exist to filter clouds from unprocessed (level 0) data [8], [9] with varying success, no method successfully removes all clouds while retaining all valid data. A secondary (postprocessing) cloud filter has been proposed by Hu et al. [10] (hereafter termed “automated filter”) which is being implemented to autonomously remove cloudManuscript received October 28, 2010; revised December 21, 2010 and January 19, 2011; accepted January 25, 2011. Date of publication March 9, 2011; date of current version June 24, 2011. This work was supported by the National Aeronautics and Space Administration under Grants NNH10AN14I (through Florida Fish and Wildlife Conservation Commission Subcontract 09249) and NNX09AV24G to the University of South Florida. The authors are with the College of Marine Science, University of South Florida, St. Petersburg, FL 33701 USA (e-mail: [email protected]; [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2011.2111353
contaminated pixels in SST images at the University of South Florida (USF). Inputs for this filter are images that have been calibrated, navigated, cloud filtered, and processed to SST using TeraScan software (version 3.2; SeaSpace Corporation) as described by Hu et al. [10] (termed “raw” images). The automated filter relies on the temporal stability and climatological regularity of SST for a given location (image pixel)—removing pixels which differ by either 2◦ from a three-day moving median value or 4◦ from the weekly climatological mean [10]. These threshold values were determined using long-term in situ temperature data collected along the Florida Keys Reef Tract, and the results were validated for the same region. An extreme cold event affected South Florida and the Gulf of Mexico in January 2010. This event revealed that, for shallow waters close to land, valid SST data beyond the climatological threshold were mistakenly discarded as clouds by the automated filter. As the cloud filter was not designed to retain anomalies larger than 4 ◦ C, other cold events in the past might also have been improperly masked for nearshore waters. An unintended consequence of this artifact is the generation of a positively biased SST climatology. The problem of false cloud masking would thus be exacerbated in filtering of future images. All SST cloud-masking approaches seek to maximize cloud exclusion while retaining valid SST observations. The failure of the automated filter to detect the January 2010 cold event required a manual retrospective assessment of the filter performance and subsequent analysis of the climatology errors. Accurate SST assessment is crucial, as physiological deficiencies and even mortality have been reported in response to cold SST for a variety of resident organisms including corals [11], [12], manatees [13], [14], sea turtles [15], mangroves [16], [17], and fishes [16], [18]. Apart from being critical to ecosystem health overall, each of these species contributes economic benefits to the State of Florida. Knowledge of extent and severity of extreme-temperature events is required to effectively monitor this coastal ecosystem and may help direct real-time research or rescue efforts. Our objective was thus to develop an improved high-resolution SST climatology to study cold events around Florida. II. M ETHODS In response to the cold event in January 2010, existing SST climatology and individual cloud-filtered images for the month of January (1995–2010) from the USF AVHRR ground station and data archive were examined. We extracted all images collected for each January from 1995 to 2010 covering waters
1545-598X/$26.00 © 2011 IEEE
770
Fig. 1. AVHRR images from the NOAA-15 satellite collected on January 15, 2010, 10:37 GMT. (a) Raw image showing cold waters off Florida’s west coast around (yellow oval) Big Bend and (red oval) cloudinduced artifacts. (b) Automated postprocessing cloud filter (red oval) accurately masks cloud pixels but (yellow oval) falsely masks cold valid SST data. (c) (White) Manually created ROI over the raw image to define to which pixels the automated filter should not be applied. (d) Final image from the hybrid postprocessing filter (combining the automated filter and manual ROI delineation).
around Florida (24◦ to 31◦ N, 79◦ to 86◦ W), representing a database of 3200 satellite images. Passes where the swath measured less than approximately 25% of the coverage area were discarded, leaving 2703 images. Raw SST images processed with TeraScan to SST but not filtered using the secondary cloud-screening algorithm of Hu et al. [10] were first visually compared side by side to corresponding postprocessing filtered images [Fig. 1(a) and (b)]. Where it was apparent that the filtered image was falsely masking valid SST data, the raw image was set aside for reprocessing. This determination of valid data was based on integration of spatial and temporal information. Specifically, improper masking was documented only when the automated cloud filter removed a region-wide (approximately 200 km2 or more) feature which persisted in three or more consecutive images and displayed smooth spatial gradients consistent with that expected from the underlying bathymetry and water flow regime. The selected raw images were loaded into IDL/ENVI (version 4.5; Research Systems, Inc.) image processing software, and a region of interest (ROI) outlining the falsely masked pixels was manually created using ENVI’s ROI tool [Fig. 1(c)]. Raw data pixels covered by the ROI were passed through the automated filter to generate a new image. Those pixels outside the ROI were superimposed on the new image without modification [Fig. 1(d)]. In the creation of the ROIs, special attention was given to coastal locations as these are more likely to show faster and stronger temperature fluctuations in response to atmospheric weather patterns. In this way, the resulting filter (hereafter termed “hybrid filter”) retained the pixels deemed
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 4, JULY 2011
Fig. 2. Differences between 1995–2010 climatologies derived from the automated filter and hybrid filter. (a) January 1–7. (b) January 8–14. (c) January 15– 21. (d) January 22–28. Negative numbers indicate that the climatology from the hybrid filter has lower SST than the climatology from the automated filter. White is zero. Areas within the yellow and red outline boxes show zoomed-in images of the Florida Bay and southern Big Bend regions, respectively.
valid from both the raw and automatically filtered images, maximizing the amount of valid data in the image. Although this method might introduce a small degree of subjectivity, interpretation from different analyzers showed excellent consistency in identifying the misclassified pixels. Note that this manual effort was not to manually delineate clouded or mixed pixels but instead only to define a crude ROI with which to override the automated filter. Two sets of monthly and weekly climatologies were created for January 1995–2010. The first was constructed using only the images processed with the automated filter (i.e., including those with false cloud-masking artifacts). The second climatology used the hybrid-filtered images and those images in the series with no artifacts detected. The difference between the two weekly climatologies was calculated by subtracting the first (automated) from the second (hybrid; Fig. 2). Finally, the reprocessed images and new climatologies were used to assess the extent and severity of the January 2010 cold event in Florida. Daily averaged SST composites from January 10 to 14, 2010, were subtracted from the weekly climatology to create anomaly maps (Fig. 3). To evaluate the results of the hybrid filter, images from both filters were compared to in situ data collected from the archived National Data Buoy Center (NDBC) sea temperature data. Of the 66 NDBC stations that measured water temperature within the region studied, 26 included data that matched the AVHRR data in both time and space (Fig. 4). Data from both sources were used only if the sample times were within 0.5 h. This led to 10 578 matching pairs for the automated filter and 10 900 matching pairs for the hybrid filter. Linear regressions, root-mean-square (rms) difference, and bias calculations were performed using IDL (version 7.0; ITT Visual Information Solutions).
BARNES et al.: IMPROVED SST CLIMATOLOGY TO ASSESS COLD WATER EVENTS OFF FLORIDA
Fig. 3. Daily composite SST anomalies on (a) and (b) January 11, (c) and (d) January 12, and (e) and (f) January 13, 2010. Images (a), (c), and (e) were created using the automated filter with the original climatology, whereas images (b), (d), and (f) were generated with the current hybrid filter and improved climatology. White pixels represent clouds or invalid data.
Fig. 4.
Map of NDBC stations used and regions discussed in this letter.
III. R ESULTS Of the 2703 images analyzed, 498 images (18.4%) showed false cloud-masking artifacts from the automated filter. The largest portion of the images that required reprocessing were
771
from extreme cold events in 2001 and 2010. The automated filter regularly masked extremely cold (i.e., differing by > 4 ◦ C from climatology) but valid SST data in nearshore waters (see Fig. 1). The incorrect masking occurred around most of Florida’s coast, particularly in the Florida Bay and Big Bend regions. Several images also showed that the automated filter occasionally incorrectly masked warm-temperature data, particularly during warm spells in January of 1997 and 2002. The weekly climatologies created with the hybrid filter showed substantially different average SST (+1 ◦ C to −2 ◦ C) than the climatologies based on the automated cloud screening. The most extreme deviations were observed nearshore during the first and second weeks of January. In the first week (January 1 to 7), the original climatology showed overall higher average SST than the hybrid climatology over the entire west coast of Florida. In the Big Bend and Florida Bay regions, this new climatology was cooler by up to 2 ◦ C. The second week of January (days 8 to 14) showed temperatures approximately 0.5 ◦ C lower in the nearshore regions but up to 1 ◦ C higher in some areas of central western Florida and in Florida Bay [Fig. 2(b)]. Neither the third (January 15 to 21) nor the fourth (January 22 to 28) week showed any large differences between the two climatologies. The daily SST composites created with the two filters were compared to their corresponding filtered monthly climatologies. The original automated filter misidentified most pixels located in waters shallower than 20 m as clouds during the extreme cold event on January 11–13, 2010 (Fig. 3). The images processed with the hybrid filter, however, show that the daily averaged SST in Florida Bay during this span had negative anomalies of up to 14 ◦ C. The maximum negative anomaly for the Big Bend region was 11.6 ◦ C. The coldest temperatures occurred on January 11, 2010, with a slight temperature increase over the next two days (Fig. 3). Unfortunately, cloud cover on January 10 and 14 precluded region-wide examination of the days immediately preceding and following these extremetemperature days. The automatically filtered AVHRR data had a strong positive relationship with near-surface SST observations from NDBC buoys (r2 = 0.896, n = 10 578, linear regression slope = 0.923, and intercept = 1.753 ◦ C; Fig. 5). The rms difference was 0.927 ◦ C, and the bias was 0.097 ◦ C. The hybrid filter resulted in more matching pairs and also higher accuracy statistics. The NDBC and hybrid-filtered AVHRR data showed a strong positive relationship (r2 = 0.911, n = 10 900, slope = 0.972, intercept = 1.323 ◦ C, rms difference = 0.925 ◦ C, and bias = 0.084 ◦ C; Fig. 5). The 333 additional matching pairs (range: 6.5 ◦ C to 27.5 ◦ C) recovered with the hybrid filter were also strongly correlated (r2 = 0.979, slope = 1.005, intercept = −0.406 ◦ C, rms difference = 0.854 ◦ C, and bias = −0.322 ◦ C). These comparisons demonstrate the high quality of the AVHRR SST estimates derived using the standard algorithms, which were improperly masked under unusually cold or warm conditions. The automated filter incorrectly omitted matching pairs below 10 ◦ C and approximately 30% of those between 10 ◦ C and 15 ◦ C and also had more than 10% misclassifications for pixels matched at temperatures > 26 ◦ C (Fig. 6).
772
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 4, JULY 2011
Fig. 5. Comparison of temperature measurements from AVHRR and NDBC sources, taken at the same location and within 0.5 h. Blue points are those captured by AVHRR images processed with the automated filter. Orange points are the additional data from the hybrid filter. Data from January 10–13, 2010, are also denoted, most of which are missed by the automated filter.
Fig. 6. Percentage of matched AVHRR/NDBC data pairs captured by the hybrid filter that were missed by the automated filter, binned in 1 ◦ C intervals. The matched AVHRR/NDBC data pairs were taken at the same location and within 0.5 h.
IV. D ISCUSSION Nearly one in five images from January 1995–2010 was improved using the hybrid filter. Such misclassifications, however, may also occur in nonwinter months. For example, improper cloud masking may occur in the wake of a hurricane, in response to unseasonal upwelling of cold water [19], or under extremely calm and warm conditions. The primary misclassifications occurred north of the Florida Keys and in the Big Bend region—shallow waters heavily influenced by rapidly fluctuating air temperatures, insolation, and wind. Colder-thanexpected water was often seen also around river outflows, owing to the mixing of cold riverine waters with marine waters in these regions.
The difference between climatologies computed using the different filters was not large, primarily because the inputs for the two climatologies were identical in 81.6% of the images and most of the pixels in the reprocessed images. Furthermore, the automated filter had improperly masked both warm and cold waters at specific locations, and these errors cancelled to some degree in the climatology calculations. However, without the hybrid approach, it would be impossible to understand and quantify either positive or negative errors. The effect can be clearly visualized in Fig. 5. The images processed with the hybrid filter helped to extend the range of the observed temperatures to ∼ 6 ◦ C and improved the matchup statistics (Figs. 5 and 6), resulting in more data coverage on individual images. Validation against NDBC data also served to justify the manual override methodology, as the additional matching pairs showed high correlation and small error. There was a minor cold bias in these data (AVHRR data colder than NDBC data) in contrast to the slight warm bias of the entire data set. The cold bias is primarily driven by a cluster of additional matching pairs between approximately 8 ◦ C and 10 ◦ C, nearly all of which are from a single location (station 41112) during the 2010 cold event. Given the high spatial heterogeneity during this time near the St. Johns River plume, these cold-biased matches possibly are artifacts of the different spatial scales of measurement (point buoy versus 1-km2 pixel satellites), or they may indicate a slight calibration error at one NDBC station. The problem of false cloud masking of cold pixels is not unique to this region [9] or to the AVHRR sensor, as it is even present in the most sophisticated MODIS observations and cloud-screening algorithms [20]. Although the visual interpretation of thousands of images appears tedious, in practice, it is feasible and represents the best approach for climatology development. The work done here manually would nevertheless be difficult to implement in an autonomous process, yet it does serve as an example in developing the most accurate high-resolution AVHRR SST climatology for coastal regions in the global ocean that may experience extreme-temperature fluctuations. An autonomous climatology-based filter which allowed all anomalies observed in this study (up to 14 ◦ C) would clearly be insufficient to remove clouds from this system. Alternatively, lowering filter thresholds for nearshore waters (for example, by applying a bathymetry mask) or during winter months could potentially create artificial “fronts” in the climatologies at these spatial or temporal boundaries, which might further affect subsequent image processing. Given the strong correlation between the AVHRR and NDBC data, these in situ sensors could potentially be integrated into the cloud-filtering methodology. Such a system might allow autonomous overrides of falsely masked areas but would require a more extensive sensor network to be synoptically effective. At present, the possibility of extreme temperatures in coastal waters of Florida (and similar coastal waters in the global ocean) highlights the need for improved autonomous cloudmasking methods for high-resolution satellite imagery. Furthermore, for effective monitoring of environmental parameters, it is important to continue coordinated efforts to capture SST and
BARNES et al.: IMPROVED SST CLIMATOLOGY TO ASSESS COLD WATER EVENTS OFF FLORIDA
bottom temperatures from local in situ sensors. Such systems allow real-time validation of filtered satellite data and provide more thorough information of the thermal conditions throughout the water column. ACKNOWLEDGMENT The authors would like to thank the staff of the University of South Florida (USF) Institute for Marine Remote Sensing (IMaRS) for their acquisition and geonavigation of the Advanced Very High Resolution Radiometer database. The authors would also like to thank the anonymous reviewers and the many contributors within IMaRS and the USF Optical Oceanography Laboratory who helped to greatly improve this letter. The National Data Buoy Center historical temperature measurements are provided by the National Oceanic and Atmospheric Administration (available at www.ndbc.noaa.gov), and the Florida coastline layer (Fig. 4) is obtained from the Florida Fish and Wildlife Research Institute. R EFERENCES [1] A. E. Strong and E. P. McClain, “Improved ocean surface temperatures from space—Comparisons with drifting buoys,” Bull. Amer. Meteorol. Soc., vol. 65, no. 2, pp. 138–142, Feb. 1984. [2] A. F. Pearce, A. J. Prata, and C. R. Manning, “Comparison of NOAA/AVHRR-2 sea surface temperatures with surface measurements in coastal waters,” Int. J. Remote Sens., vol. 10, no. 1, pp. 37–52, Jan. 1989. [3] W. G. Pichel, “Operational production of multichannel sea surface temperatures from NOAA polar satellite AVHRR data,” Global Planet. Change, vol. 4, no. 1–3, pp. 173–177, Jul. 1991. [4] C. C. Walton, W. G. Pichel, J. F. Sapper, and D. A. May, “The development and operational application of nonlinear algorithms for the measurement of sea surface temperatures with the NOAA polar-orbiting environmental satellites,” J. Geophys. Res., vol. 103, pp. 27 999–28 012, 1998. [5] X. Li, W. G. Pichel, P. Clemente-Colón, V. Krasnopolsky, and J. Sapper, “Validation of coastal sea and lake surface temperature measurements derived from NOAA/AVHRR data,” Int. J. Remote Sens., vol. 22, no. 7, pp. 1285–1303, 2001.
773
[6] E. P. McClain, W. G. Pichel, C. C. Walton, Z. Ahmad, and J. Sutton, “Multi-channel improvements to satellite-derived global sea surface temperatures,” Adv. Space Res., vol. 2, no. 6, pp. 43–47, 1983. [7] A. K. Heidinger, V. R. Anne, and C. Dean, “Using MODIS to estimate cloud contamination of the AVHRR data record,” J. Atmos. Ocean. Technol., vol. 19, no. 5, pp. 586–601, May 2002. [8] R. W. Saunders and K. T. Kriebel, “An improved method for detecting clear sky and cloudy radiances from AVHRR data,” Int. J. Remote Sens., vol. 9, no. 1, pp. 123–150, Jan. 1988. [9] C. J. Merchant, A. R. Harris, E. Maturi, and S. Maccallum, “Probabilistic physically based cloud screening of satellite infrared imagery for operational sea surface temperature retrieval,” Q. J. R. Meteorol. Soc., vol. 131, pp. 2735–2755, 2005. [10] C. Hu, F. Muller-Karger, B. Murch, D. Myhre, J. Taylor, R. Luerssen, C. Moses, C. Zhang, L. Gramer, and J. Hendee, “Building an automated integrated observing system to detect sea surface temperature anomaly events in the Florida Keys,” IEEE Trans. Geosci. Remote Sens., vol. 47, pp. 2071–2084, Jul. 2009. [11] P. Jokiel and S. Coles, “Effects of temperature on the mortality and growth of Hawaiian reef corals,” Mar. Biol., vol. 43, no. 3, pp. 201–208, Sep. 1977. [12] A. E. Douglas, “Coral bleaching—How and why?” Mar. Pollut. Bull., vol. 46, no. 4, pp. 385–392, Apr. 2003. [13] A. B. Irvine, “Manatee metabolism and its influence on distribution in Florida,” Biol. Conserv., vol. 25, no. 4, pp. 315–334, Apr. 1983. [14] C. J. Deutsch, J. P. Reid, R. K. Bonde, D. E. Easton, H. I. Kochman, and T. J. O’Shea, “Seasonal movements, migratory behavior, and site fidelity of west Indian manatees along the Atlantic coast of the United States,” Wildlife Monogr., vol. 151, pp. 1–77, 2003. [15] B. E. Witherington and L. M. Ehrhart, “Hypothermic stunning and mortality of marine turtles in the Indian river lagoon system, Florida,” Copeia, vol. 1989, pp. 696–703, 1989. [16] M. Storey and E. W. Gudger, “Mortality of fishes due to cold at Sanibel Island, Florida 1886–1936,” Ecology, vol. 17, no. 4, pp. 640–648, Oct. 1936. [17] T. Savage, “Florida mangroves as shoreline stabilizers,” Florida Dept. Natural Resour., St. Petersburg, FL, Prof. Pap. No. 19, 1972, 46 pp. [18] Y. Sadovy and A.-M. Eklund, “Synopsis of Biological Data on the Nassau Grouper, Epinephelus Striatus (Bloch, 1792), and the Jewfish, E. Itajara (Lichtenstein, 1822),” Seattle, WA, U.S. Dept. Commerce, Nat. Ocean. Atmos. Admin., Nat. Mar. Fisheries Service, 65 pp., 1999. [19] N. D. Walker, R. R. Leben, and S. Balasubramanian, “Hurricane-forced upwelling and chlorophyll a enhancement within cold-core cyclones in the Gulf of Mexico,” Geophys. Res. Lett., vol. 32, p. L18 610, Sep. 2005. [20] S. A. Ackerman, K. I. Strabala, W. P. Menzel, R. A. Frey, C. C. Moeller, and L. E. Gumley, “Discriminating clear sky from clouds with MODIS,” J. Geophys. Res.—Atmos., vol. 103, no. D24, pp. 32 141–32 157, Dec. 1998.
